Deep Learning Approach to Multi-phenomenological Nuclear Fuel Cycle Signals for Nonproliferation Applications

In order to reduce the time required for data analysis and decision-making relevant to nuclear proliferation detection, Artificial Intelligence (AI) techniques are applied to multi-phenomenological signals emitted from nuclear fuel cycle facilities to identify non-human readable characteristic signatures of operations for use in detecting proliferation activities. Seismic and magnetic emanations were collected in the vicinity of the High Flux Isotope Reactor (HFIR) and the McClellan Nuclear Research Center (MNRC). A novel bi-phenomenology DL network is designed to test the viability of transfer learning between nuclear reactor facilities. It is found that the network produces an 84.1% accuracy (99.4% without transient states) for predicting the operational state of the MNRC reactor when trained on the operational state of the HFIR reactor. In comparison, the best performing traditional ML single-phenomenology algorithm, K-Means, produces a 67.8% prediction accuracy (80.5% without transient states)